A server for CodeFuse-CGM, a graph-integrated large language model designed for repository-level software engineering tasks.
A Model Context Protocol (MCP) server implementation of CodeFuse-CGM (Code Graph Model), providing graph-integrated large language model capabilities for repository-level software engineering tasks.
# Clone the repository
git clone https://github.com/your-org/cgm-mcp.git
cd cgm-mcp
# Install dependencies
pip install -r requirements.txt
# Or install in development mode
pip install -e .
# No additional setup needed!
# MPS (Metal Performance Shaders) is automatically detected and enabled
pip install torch torchvision torchaudio # Usually already installed
# Install CUDA-enabled PyTorch
pip install torch --index-url https://download.pytorch.org/whl/cu118
# Optional: Enhanced GPU features
pip install cupy-cuda11x # For CUDA 11.x
# or
pip install cupy-cuda12x # For CUDA 12.x
# Linux (ROCm)
pip install torch --index-url https://download.pytorch.org/whl/rocm5.6
# Windows (DirectML)
pip install torch-directml
pip install cgm-mcp
# Run setup script
./scripts/setup.sh
# Copy example environment file
cp .env.example .env
# Edit .env with your API keys
export CGM_LLM_PROVIDER=openai
export CGM_LLM_API_KEY=your-openai-api-key
export CGM_LLM_MODEL=gpt-4
# Install and start Ollama
curl -fsSL https://ollama.ai/install.sh | sh
ollama serve
# Download recommended model
ollama pull deepseek-coder:6.7b
# Start with local model
./scripts/start_local.sh --provider ollama --model deepseek-coder:6.7b
# Download and start LM Studio
# Load deepseek-coder-6.7b-instruct model
# Start local server
# Start CGM with LM Studio
./scripts/start_local.sh --provider lmstudio
# Start MCP server (cloud models)
python main.py
# Start with local models
./scripts/start_local.sh
# Or with custom config
python main.py --config config.local.json --log-level DEBUG
# Run example usage
python examples/example_usage.py
# Check GPU acceleration status
python check_gpu_dependencies.py
# Run the GPU dependency checker
python check_gpu_dependencies.py
Expected output for Apple Silicon:
🎉 OPTIMAL: Apple Silicon GPU acceleration is active!
• MPS backend enabled
• No additional dependencies needed
• CuPy warnings can be ignored
| Platform | Entity Matching | Text Processing | Cache Hit Rate |
|---|---|---|---|
| Apple Silicon (MPS) | 42x speedup (cached) | ~0.001s (200 files) | 95%+ |
| NVIDIA CUDA | 5-10x speedup | 3-5x speedup | 90%+ |
| AMD ROCm | 3-8x speedup | 2-4x speedup | 90%+ |
| CPU Fallback | Baseline | Baseline | 85%+ |
# GPU Configuration
export CGM_USE_GPU=true # Enable GPU acceleration
export CGM_GPU_BATCH_SIZE=1024 # Batch size for GPU operations
export CGM_SIMILARITY_THRESHOLD=0.1 # Entity similarity threshold
export CGM_CACHE_EMBEDDINGS=true # Cache embedding vectors
# Memory Management
export CGM_MAX_CACHE_SIZE=500 # Maximum cached files
export CGM_MEMORY_CLEANUP_THRESHOLD=80 # Memory cleanup trigger (%)
export CGM_GPU_MEMORY_FRACTION=0.8 # GPU memory usage limit
{
"gpu": {
"use_gpu": true,
"batch_size": 1024,
"max_sequence_length": 512,
"similarity_threshold": 0.1,
"cache_embeddings": true,
"gpu_memory_fraction": 0.8
},
"performance": {
"max_concurrent_files": 10,
"cache_ttl_seconds": 3600,
"max_file_cache_size": 500,
"memory_cleanup_threshold": 80
}
}
| Variable | Description | Default |
|---|---|---|
CGM_LLM_PROVIDER | LLM provider (openai, anthropic, ollama, lmstudio, mock) | openai |
CGM_LLM_API_KEY | API key for LLM provider (not needed for local models) | Required for cloud |
CGM_LLM_MODEL | Model name | gpt-4 |
CGM_LLM_API_BASE | Custom API base URL (for local models) | Provider default |
CGM_LLM_TEMPERATURE | Generation temperature | 0.1 |
CGM_LOG_LEVEL | Logging level | INFO |
Create a config.json file:
{
"llm": {
"provider": "openai",
"model": "gpt-4",
"temperature": 0.1,
"max_tokens": 4000
}
}
{
"llm": {
"provider": "ollama",
"model": "deepseek-coder:6.7b",
"api_base": "http://localhost:11434",
"temperature": 0.1,
"max_tokens": 4000
},
"graph": {
"max_nodes": 5000,
"max_edges": 25000,
"cache_enabled": true
},
"server": {
"log_level": "INFO",
"max_concurrent_tasks": 3
}
}
The server provides the following MCP tools:
cgm_analyze_repositoryAnalyze repository structure and extract code entities with GPU acceleration.
Parameters:
repository_path: Path to the repositoryquery: Search query for relevant codeanalysis_scope: Scope of analysis (full, focused, minimal)max_files: Maximum number of files to analyzecgm_get_file_contentGet detailed file content and analysis with concurrent processing.
Parameters:
repository_path: Path to the repositoryfile_paths: List of file paths to analyzecgm_find_related_codeFind code entities related to a specific entity using GPU-accelerated similarity matching.
Parameters:
repository_path: Path to the repositoryentity_name: Name of the entity to find relations forrelation_types: Types of relations to include (optional)cgm_extract_contextExtract structured context for external model consumption.
Parameters:
repository_path: Path to the repositoryquery: Query for context extractionformat: Output format (structured, markdown, prompt)clear_gpu_cacheClear GPU caches to free memory.
Parameters: None
cgm_process_issueProcess a repository issue using the CGM framework.
Parameters:
task_type: Type of task (issue_resolution, code_analysis, bug_fixing, feature_implementation)repository_name: Name of the repositoryissue_description: Description of the issuerepository_context: Optional repository contextExample:
{
"task_type": "issue_resolution",
"repository_name": "my-project",
"issue_description": "Authentication fails with special characters in password",
"repository_context": {
"path": "/path/to/repository",
"language": "Python",
"framework": "Django"
}
}
cgm_get_task_statusGet the status of a running task.
Parameters:
task_id: Task ID to checkcgm_health_checkCheck server health status.
cgm://health: Server health informationcgm://tasks: List of active taskscgm://cache: Cache statistics and hit/miss ratioscgm://performance: Server performance and memory usage metricscgm://gpu: GPU acceleration status and memory usage# Check GPU status
curl "cgm://gpu"
# Monitor cache performance
curl "cgm://cache"
# View performance metrics
curl "cgm://performance"
CGM follows a four-stage pipeline:
graph LR
A[Issue] --> B[Rewriter]
B --> C[Retriever]
C --> D[Reranker]
D --> E[Reader]
E --> F[Code Patches]
G[Code Graph] --> C
G --> D
G --> E
Constructs repository-level code graphs by analyzing:
CGMRequestclass CGMRequest(BaseModel):
task_type: TaskType
repository_name: str
issue_description: str
repository_context: Optional[Dict[str, Any]] = None
CGMResponseclass CGMResponse(BaseModel):
task_id: str
status: str
rewriter_result: Optional[RewriterResponse]
retriever_result: Optional[RetrieverResponse]
reranker_result: Optional[RerankerResponse]
reader_result: Optional[ReaderResponse]
processing_time: float
CodePatchclass CodePatch(BaseModel):
file_path: str
original_code: str
modified_code: str
line_start: int
line_end: int
explanation: str
import asyncio
from cgm_mcp.server import CGMServer
from cgm_mcp.models import CGMRequest, TaskType
async def resolve_issue():
server = CGMServer(config)
request = CGMRequest(
task_type=TaskType.ISSUE_RESOLUTION,
repository_name="my-app",
issue_description="Login fails with special characters",
repository_context={"path": "./my-app"}
)
response = await server._process_issue(request.dict())
for patch in response.reader_result.patches:
print(f"File: {patch.file_path}")
print(f"Changes: {patch.explanation}")
Add to your Claude Desktop MCP configuration:
{
"mcpServers": {
"cgm": {
"command": "python",
"args": ["/path/to/cgm-mcp/main.py"],
"env": {
"CGM_LLM_API_KEY": "your-api-key"
}
}
}
}
{
"memory": {
"gpu_available": true,
"platform": "Apple Silicon",
"backend": "Metal Performance Shaders",
"gpu_memory_allocated": 0.6
},
"performance": {
"gpu_entity_matches": 15,
"cache_hit_rate": 94.2
}
}
{
"analysis_cache": {"size": 45, "maxsize": 100},
"file_cache": {"size": 234, "maxsize": 500},
"stats": {"hits": 156, "misses": 23, "hit_rate": 87.2}
}
# Check GPU acceleration status
python check_gpu_dependencies.py
# Run performance tests
python gpu_verification.py
python test_multiplatform_gpu.py
# Run tests
pytest tests/
# Run with coverage
pytest tests/ --cov=cgm_mcp
# Run specific test
pytest tests/test_components.py::TestRewriterComponent
# Performance tests
python test_gpu_acceleration.py
python gpu_verification.py
cgm://gpu, cgm://cache, cgm://performancegit checkout -b feature/amazing-feature)git commit -m 'Add amazing feature')git push origin feature/amazing-feature)This project is licensed under the MIT License - see the LICENSE file for details.
CGM MCP Server - Bringing graph-integrated code intelligence to your development workflow! 🚀
🍎 MCP server for Xcode's xctrace, xcrun, xcodebuild.
Open-source tool for collaborative editing, versioning, evaluating, and releasing prompts.
Interact with the JFrog Platform API for repository management, build tracking, and release lifecycle management.
A server for Zep, a long-term memory store for AI applications, requiring a ZEP_API_KEY for access.
Make your AI agent speak every language on the planet, using Lingo.dev Localization Engine.
MCP server for Autodesk Maya
MCP Expr-Lang provides a seamless integration between Claude AI and the powerful expr-lang expression evaluation engine.
Tools to query latest Maven dependency information
Turns any command-line interface (CLI) command into a simple StdIO-based MCP server.
A Python server providing Retrieval-Augmented Generation (RAG) functionality. It indexes various document formats and requires a PostgreSQL database with pgvector.